[AIR] Add TorchDetectionPredictor

[bveeramani]

Description

Add a class TorchDetectionPredictor that works with TorchVision detection models.

Use case

TorchPredictor doesn't work with TorchVision detection models because they return lists of dictionaries. So, you need to extend TorchPredictor:

class CustomTorchPredictor(TorchPredictor):
    def _predict_numpy(
        self, data: np.ndarray, dtype: torch.dtype
    ) -> Dict[str, np.ndarray]:
        inputs = [torch.as_tensor(image) for image in data["image"]]
        assert all(image.dim() == 3 for image in inputs)
        outputs = self.call_model(inputs)
        predictions = collections.defaultdict(list)
        for output in outputs:
            for key, value in output.items():
                predictions[key].append(value.cpu().detach().numpy())
        for key, value in predictions.items():
            predictions[key] = _create_possibly_ragged_ndarray(value)
        predictions = {"pred_" + key: value for key, value in predictions.items()}
        return predictions